Resolves #13946

Native Ollama embedding now uses a fail-loud request shape: /api/embed sends truncate:false, and TextEmbeddingService passes AiConfig.localModels.embedding.contextLimitTokens as native options.num_ctx for both single and batch Ollama embeddings. This keeps KB / Memory Core from accepting silently truncated vectors when deployments route embeddings through native Ollama.

Evidence: L2 focused unit/static evidence -> L2 required for the #13946 request-shape ACs. Residual: none for the close target.

Deltas from ticket

The provider now defaults native embedding truncation to false at OllamaProvider.embed() itself, so direct provider consumers inherit the same fail-loud posture. TextEmbeddingService supplies the role-specific context leaf at its native-Ollama use sites, per ADR 0019.

Test Evidence

• npm run agent-preflight -- ai/provider/Ollama.mjs ai/services/memory-core/TextEmbeddingService.mjs test/playwright/unit/ai/provider/KeepAlive.spec.mjs test/playwright/unit/ai/services/memory-core/TextEmbeddingService.spec.mjs passed.
• npm run test-unit -- test/playwright/unit/ai/provider/KeepAlive.spec.mjs test/playwright/unit/ai/services/memory-core/TextEmbeddingService.spec.mjs passed: 30 passed, 1 skipped.
• git diff --check passed.

Post-Merge Validation

• Native-Ollama deployment smoke can embed an over-context fixture and observe provider rejection instead of silent truncation.

Commits

• feb2b2c9cd — fail loud on native Ollama embedding truncation

Authored by Euclid (GPT-5, Codex Desktop). Session cd2b88d7-8134-4ef8-a10f-0b1e80c03db4.